Image Similarity Calculation And GPU Acceleration Research

Image similarity is widely used in CBIR, image matching, image recognition and machine vision. Most of the image similarity researches are included in CBIR, image matching, image recognition and image mosaic. Image similarity is a general concept, both the high-level semantic similarity, and low-level visual feature’s similarity, different people has different understand. We sum up six classes of objective image similarity, and discuss how to compute similarity of these six classes. There are many features can be used to compute image similarity. We mainly using color, texture, shape and local invariant image features. After feature extracted, use image hash or image perceptual hash to generate image hash, and use Hamming distance to measure image similarity. At last, use GPGPU to accelerate the algorithms of this paper. The main content of this dissertation can be summarized as follows:1. We use color, texture and shape features to compute image similarity, use the color histogram, color moment, and refer to the descriptors of MPEG-7,such as color layout descriptor, the main color descriptor and color structure descriptor. We simplify these descriptors. We use gray level co-occurrence matrix and Gabor wavelet transform to extract texture feature. We use Hu moment to calculate shape feature. Introduce the principle of measures algorithms in detail, and discuss which measure fit which image feature.2. We use both Harris and SIFT two classic image local invariant features and HOG algorithm to measure image similarity, and analyze the affect of the degrees scale which decide a orient.3. Combining image hash techniques, we use three image feature extraction algorithms to generate the image hash. First, use a simple image scale compression and binary. Second, extract low frequency coefficients after DCT transform. Third, use SIFT feature vector. At last, use Hamming distance to measure image similarity.4. According the above researches, we design and achieve an image similarity measure system, including the calculation of image features, similarity measure and similar image search. Using GPU to accelerate feature extract and distance metric which contain parallel computing. Experiment shows the system is reliable, stable and scalable.